Understanding how spatial variation is linked to diversity maintenance in natural communities is a pillar of plant community ecology. Theoretically, a variable landscape can maintain diversity via niche partitioning: different species can trade off in performing better or worse depending on the conditions of the patch they are growing in, and as a result, more species can sustainably coexist in a community than if it were spatially heterogeneous. In the hyperdiverse system of native annual plants in Western Australia, fallen logs may be one of the greatest contributors to generating spatial variation that could help maintain species diversity. Considerable anecdotal evidence suggests that fallen logs generate spatial variation, or patchiness, in the environment (Figure 1), and that species or assemblages of plants may respond differently depending on if they are near logs or not. Despite such anecdotal evidence, it is yet unknown if and how fallen logs contribute to maintaining species diversity in the native annual plant communities of the Western Australian wheat belt.

Figure 1: image of annual plant halos around logs
Figure 1: image of annual plant halos around logs

The project will address the following questions:

Q1) Are/how are plant communities in fallen log patches different from patches that are in the open?

Q2) Why are plant communities in fallen log patches different from patches in the open?

Q3) Are/how are plant species performances affected by proximity to fallen logs?

Hypotheses

The null hypothesis, H0, is that annual plants in fallen log patches are not different in diversity, abundance, or composition from open patches.

In addition to the null hypothesis, the following constitute four, non-mutually exclusive hypotheses concerning how fallen logs may introduce spatial variation in the environment. I include corresponding predictions for how plant communities may differ between fallen log patches as compared to open patches.

H1: Log decomposition creates islands of fertility directly around the fallen log.
Prediction 1: Nutrient composition around logs will be higher than in open plots

Prediction 2: Variations in nutrient composition in log vs open environments will correspond to variations in species composition, abundance, and/or richness in these environments.

Prediction 3: All sown plants will perform best in environments where organic logs have been left ‘insitu’. In locations where logs have been removed or replaced with pvc, the legacy of the nutrient island effect will yield higher sown plant performance than when compared to locations where logs have never been. The effect of the nutrient island in locations where logs have been added to open environments should yeild higher plant performance over time. note: performance is measured in terms of germination rate, survival to fruiting, fecundity, and/or biomass.

H2: Fallen logs alter the microclimate directly around them by providing shade.
Prediction 1: Shade and temperature around logs vs in open plots will be different

Prediction 2: Variation in shade and temperature in log vs open environments will correspond to variation in species composition, abundance, and/or richness in these environments.

Prediction 3: All sown plants will perform best in environments where there are organic or pvc logs, no matter if they have been recently moved or not.

H3: Fallen logs trap dispersing seeds as they are blown along the ground.

Prediction 1: Dispersing seeds accumulate around logs, leading to a denser stand of plants in fallen log patches. Plant abundance in fallen log patches will be higher as compared to open patches. Rare plants will be more common in fallen log patches as compared to open patches

Prediction 2: All sown plants will perform the same in all experimental environments

H4: At least some species perform differently according to variation in log vs. open environments and have short dispersal kernels, causing fitness-density covariance

Figure 2: Photo before germination, after a rain. Notice the seeming wet halo under and around the branch
Figure 2: Photo before germination, after a rain. Notice the seeming wet halo under and around the branch


Experimental Design

In this experiment, 224 plots are arranged in 7 blocks of 32 plots each within the Caron Dam nature reserve. A map can be found here.
note: the location info for 3.02 is probably incorrect as of May 2022, and location info is currently unavailable for plots 6.25 and 7.19

Each block is approximately 30m X 30m in area. Plots are 1m long and linear, and have a pin tag on either end (see Figure 3). The pin tags have the identity of the plot written on them in the form of “blocknumber.plotnumber”. Plots are 1m or more away from each other.

In each block, plot environments can be one of six types:
- A 1m log that is out in the open (open_with_log, 4 plots)
- A 1m log that is a part of a tree (insitu_log, 4 plots)
- A 1m pvc pipe that is out in the open (open_with_pvc, 4 plots)
- A 1m pvc pipe that is a part of a tree (insitu_pvc, 4 plots)
- A plot that is out in the open (open, 8 plots)
- A gap in a log where a log used to be (gap, 8 plots)

In half of the plots (not including open plots), the addition, exchange, or removal of logs or pvc to the environment was implemented in October 2020, before seed dispersal. In the other half of these plots, these manipulations were implemented after seed dispersal, in March 2021.

Within each 1m long plot, there is a ~20cm long microtransect. The ends of the microtransects are marked by a nail and a washer sunken into the ground. Each microtransect is approximately 21 cm in internal length from inner washer edge to inner washer edge. Microtransects are not sided.

In half of all plots, seeds were sown in March 2021 and February 2022. In these plots, 15 seeds each of Trachymene ornata (TROR), Goodenia rosea (GORO), and Trachymene cyanopetala (TRCY) are sown outside of the microtransects as in the diagram. These plants were selected because they represent plants common to communities next to logs (TROR), out in the open (GORO), or both (TRCY). The plots where seeds were sown are called ‘lambda’ plots as noted in Figure 3. In the dataset, the rows with a “1” in the ‘seeding_trt’ column are the plots that had seeds sown into them.

Figure 3: plot schematic
Figure 3: plot schematic

Datasets

The sets of data that we have collected for this arm of the project are the following.

  1. Community data, before and after the experiment was implemented.
  1. Soil nutrient analysis in the open and insitu log plots.
  1. Performance data of TROR, TRCY, and TROR.

Methods and analysis

Q1 Are/how are plant communities in fallen log patches different from patches that are in the open?

Overview of results

Quick recap:

  • Plant abundance did not differ between log and open patches every year.
  • Plant diversity is higher in log patches in 2022. When pulling data from 2020-2022 together, plant diversity and likelihood of having positive Shannon diversity index are higher in log patches.
  • Plot type (log vs open) explained a relatively small portion of the variance in plant species composition across plots ranging from 6-12% across years. Visually, the composition of plants tends to consistently differ between plot types in the non-dimensional space.

To compare the characteristic of plant communities in fallen log patches and open patches, we assessed differences in plant abundance, diversity using the Shannon diversity index, and plant composition summarized at the transect level. The analysis is organized by year - we first analysed plot level community data that was compiled from 2020 to 2022. In other words, each row in the data set represents an individual plot from a specific year. After that, we analysed plot level data from 2020, 2021, and 2022 separately. Next, we analysed the plot level data for each year separately: 2020, 2021, and 2022. This resulted in four levels of analysis: 1. 2020-2022 composite data, 2. 2020 data, 3. 2021 data, and 4. 2022 data. For each year, we evaluated three response variables: abundance, diversity, and plant composition.

2020-2022 There are three response variables: count (Poisson), Shannon diversity index (hurdle: binomial and truncated-Gaussian distributiobns), and composition matrix.

  1. Abundance – the abundance of plants in log patches and open patches that are similar (treatment: p > 0.05). The presented model has a structure of “count ~ treatment(log/open) + (1|block) + (1|year)”.
  2. Diversity – the probability of having a positive Shannon diversity index in the plant community is lower in open patches by 6.04% (treatment: p = 0.03). The plant diversity, as measured by the Shannon diversity index, is higher in log patches by 0.09 units (treatment: p = 0.02). The presented model has a structure of “diversity ~ treatment(log/open) + (1|block) + (1|year)”.
  3. Composition – the five-dimensional NMDS ordination has stress values of 9%, which suggests a fairly good fit for the ordinations. Each two-dimensional NMDS plot can be separated into two groups that overlap - fallen log patches and open patches. In the nondimensional space, the relative position of log plots tend to consistently deviate away from open plots. This can be interpreted as a turnover in species composition at the scale of the plot type. At the same time, relative position of plots sampled from the same block are typically close to each other in nondimensional space, reflecting the similarity of plant communities at a greater spatial level (i.e. block).

Adjusted R2 of the partial RDA is 5.62%. The conditioned terms Block and Year accounted for 44.3% of variations in species composition. Plot type (constrained term) accounted for about 6% of variations in species composition. 49.6% of variation is unexplained (unconstrained).

2020

There are three response variables: count (Poisson), Shannon diversity index (hurdle: binomial and Gaussian), and composition matrix.

  1. Abundance – the abundance of plants in log patches and open patches that are similar (treatment: p > 0.05). The presented model has a structure of “count ~ treatment(log/open) + (1|block)”.
  2. Diversity – the plant diversity and the probability of having a positive Shannon diversity index in the plant community between log and open patches are similar (treatment: p > 0.05). The presented model has a structure of “diversity ~ treatment(log/open) + (1|block)”.
  3. Composition – the four-dimensional NMDS ordination has stress values of 4.8%, which suggests a very good fit for the ordinations. Similar discussion as in 2020-2022.

Adjusted R2 of the partial RDA is 8.5%. The conditioned term, Block, accounted for 69.6% of variations in species composition. Plot type (constrained term) accounted for about 8.3% of variations in species composition. 22.6% of variation is unexplained (unconstrained).

2021

There are three response variables: count (Poisson), Shannon diversity index (hurdle: binomial and Gaussian), and composition matrix.

  1. Abundance – the abundance of plants in log patches and open patches that are similar (treatment: p > 0.05). The presented model has a structure of “count ~ treatment(log/open) + (1|block)”.
  2. Diversity – the plant diversity and the probability of having a positive Shannon diversity index in the plant community between log and open patches are similar (treatment: p > 0.05). The presented model has a structure of “diversity ~ treatment(log/open) + (1|block)”.
  3. Composition – The five-dimensional NMDS ordination has stress values of 4.1%, which suggests a very good fit for the ordinations. Similar discussion as in 2020-2022.

Adjusted R2 of the partial RDA is 11.9%. The conditioned term, Block, accounted for 53.7% of variations in species composition. Plot type (constrained term) accounted for about 12.1% of variations in species composition. 34.2% of variation is unexplained (unconstrained).

2022

There are three response variables: count (Poisson), Shannon diversity index (hurdle: binomial and Gaussian), and composition matrix.

  1. Abundance – the abundance of plants in log patches and open patches that are similar (treatment: p > 0.05). The presented model has a structure of “count ~ treatment(log/open) + (1|block)”.
  2. Diversity – the probability of having a positive Shannon diversity index in the plant community between log and open patches are similar (treatment: p > 0.05). The plant diversity, as measured by the Shannon diversity index, higher in log patches by 0.24 units (treatment: p = 0.007). The presented model has a structure of “diversity ~ treatment(log/open) + (1|block)”.
  3. Composition – The five-dimensional NMDS ordination has stress values of 4%, which suggests a very good fit for the ordinations. Similar discussion as in 2020-2022.

Adjusted R2 of the partial RDA is 11.0%. The conditioned term, Block, accounted for 55.7% of variations in species composition. Plot type (constrained term) accounted for about 11.4% of variations in species composition. 33.0% of variation is unexplained (unconstrained).

Statistical Methods

  1. Abundance comparison

Comparison of plant abundance between log patches and open patches involves two types of data: (1) plant counts of all plots in 2020 before the experiment setup differentiated by the initial treatment “log” and “open”; and (2) plant counts from insitu_log and insitu_open plots in 2021 and 2022 respectively. All count data used for abundance comparison includes species with unknown identity. We used a generalized linear mixed-effects model (GLMER) to model the abundance of plant individuals (2020, 2021 and 2022 data) in unaltered fallen log patches and open patches using a Poisson distribution structure with block as random terms. When pulling count data from 2020 to 2022 together, we included time as a random term to avoid pseudo-replication.

  1. Diversity comparison

A 20-centimeter linear transect was set up for each plot in 2020. We recorded species identity and individual counts for all species occurring along the transects in 2020, 2021 and 2022 respectively. Comparison of plant diversity between log patches and open patches involves two types of data: (1) plant count for each species of all plots in 2020 before the experiment setup differentiated by the initial treatment “log” and “open”; and (2) plant count for each species from insitu_log and insitu_open plots in 2021 and 2022 respectively. All data used for diversity comparison excludes species with unknown identity.

We used the Shannon diversity index to calculate the plant species diversity for each transect at the plot level. A considerable number of zeros are generated in the computation of the Shannon diversity index. We first tried modelling diversity data with a linear mixed-effect model (LMER) assuming a Gaussian distribution. The resulting model was zero-inflated as testing with DHARMa.

As a solution, we used a two-component hurdle model on the truncated data set where the response diversity index was separated into zero and non-zero values. In the first component, we used a binomial GLMER to model the probability of zeros and non-zeros in the diversity index with blocks as random terms. For the second component, we assumed an approximate Gaussian distribution of the non-zero Shannon index and used a linear mixed-effects regression (LMER) to model the response non-zero diversity index with treatment as a fixed effect (log vs open) and block as a random factor. Specifically, for the hurdle model on composite diversity data (2020 to 2022), we also included year as a random factor.

  1. Composition

Comparison of plant composition between log patches and open patches involves two types of data: (1) plant count for each species of all plots in 2020 before the experiment setup differentiated by the initial treatment “log” and “open”; and (2) plant count for each species from insitu_log and insitu_open plots in 2021 and 2022 respectively. All data used for diversity comparison excludes species with unknown identity. For transects where no plants were found, an artificial species, “x,” was added to the species composition matrix to represent zero plants.

We used Non-Metric Dimensional Scaling (NMDS) to analyse the differences in plant communities with Bray-Curtis dissimilarity metrics. We retained two dimensions in the NMDS ordination plot for the visualization and interpretation of plant community composition differences between plot types and blocks.

To identify the contribution of each species to the compositional changes amongst different sample plots, we extracted ordination scores for each species along NMDS axes 1 and 2. The species scores represent the weighted average of a species’ abundance score in a sample community along our selected NMDS axes. We also ran a partial Redundancy Analysis (RDA) to test the correlation between species composition (squared-root transformed) and plot type (log vs open patch) with block as a random term.

Analysis

Abundance analysis
** 2020 - 2022 **
  Abundance:2020-2022 Abundance:2020 Abundance:2021 Abundance:2022
Predictors Incidence Rate Ratios CI p Incidence Rate Ratios CI p Incidence Rate Ratios CI p Incidence Rate Ratios CI p
(Intercept) 6.79 4.88 – 9.46 <0.001 4.74 4.34 – 5.18 <0.001 7.14 6.20 – 8.22 <0.001 8.82 7.56 – 10.30 <0.001
init [open] 1.00 0.93 – 1.09 0.913 0.94 0.83 – 1.07 0.355 1.09 0.92 – 1.28 0.322 1.04 0.89 – 1.21 0.642
Random Effects
σ2 0.14 0.20 0.12 0.10
τ00 0.00 block 0.00 block 0.00 block 0.01 block
0.08 time      
ICC 0.38 0.00 0.03 0.12
N 7 block 7 block 7 block 7 block
3 time      
Observations 391 221 86 84
Marginal R2 / Conditional R2 0.000 / 0.378 0.004 / 0.009 0.012 / 0.037 0.002 / 0.127

Diversity analysis
Shannon diversity index (SDI): 2020-2022
  SDI [all years]
Predictors Estimates CI p
Count Model
(Intercept) 0.89 0.81 – 0.98 <0.001
init [open] -0.13 -0.23 – -0.03 0.013
(Intercept) 0.50 0.47 – 0.54
Zero-Inflated Model
(Intercept) -20.83 -10177.83 – 10136.17 0.997
Random Effects
σ2 0.50
τ00 block 0.00
ICC 0.01
N block 7
Observations 393
Marginal R2 / Conditional R2 0.008 / 0.013
Shannon diversity index (SDI): 2020
  SDI [2020]
Predictors Estimates CI p
Count Model
(Intercept) 0.78 0.68 – 0.88 <0.001
init [open] -0.13 -0.26 – 0.01 0.063
(Intercept) 0.51 0.46 – 0.56
Zero-Inflated Model
(Intercept) -20.98 -13594.28 – 13552.31 0.998
Random Effects
σ2 0.51
τ00 block 0.00
ICC 0.01
N block 7
Observations 223
Marginal R2 / Conditional R2 0.008 / 0.013
Shannon diversity index (SDI): 2021
  SDI [2021]
Predictors Estimates CI p
Count Model
(Intercept) 1.06 0.92 – 1.19 <0.001
init [open] -0.02 -0.20 – 0.16 0.822
(Intercept) 0.36 0.28 – 0.45
Zero-Inflated Model
(Intercept) -2.47 -3.46 – -1.48 <0.001
Random Effects
σ2 0.36
τ00 block 0.00
N block 7
Observations 86
Marginal R2 / Conditional R2 0.000 / NA
Shannon diversity index (SDI): 2022
  SDI [2022]
Predictors Estimates CI p
Count Model
(Intercept) 1.21 1.03 – 1.40 <0.001
init [open] -0.38 -0.57 – -0.18 <0.001
(Intercept) 0.42 0.36 – 0.50
Zero-Inflated Model
(Intercept) -22.39 -24824.98 – 24780.19 0.999
Random Effects
σ2 0.42
τ00 block 0.02
ICC 0.04
N block 7
Observations 84
Marginal R2 / Conditional R2 0.068 / 0.105

Composition analysis

Composition dissimilarity * 2020 *

Partitioning_of_variance Inertia Proportion Note
Call:rda(X = ass.rel.t0, Y = init, Z = block) Model signif. = *
Total 0.42624 1 R2 = 0.08471055
Conditioned 0.29429 0.69045 Variance explained by ‘block’
Constrained 0.03551 0.08332 Variance explained by ’Treatment
Unconstrained 0.09643 0.22624 Unexplained variance

Composition dissimilarity * 2021 *

Partitioning_of_variance Inertia Proportion Note
Call:rda(X = ass.rel.t1, Y = init, Z = block) Model signif. = .
Total 0.53262 1 R2 = 0.1186067
Conditioned 0.28621 0.5373 Variance explained by ‘block’
Constrained 0.06436 0.1208 Variance explained by ’Treatment
Unconstrained 0.18206 0.3418 Unexplained variance

Composition dissimilarity * 2022 *

Partitioning_of_variance Inertia Proportion Note
Call:rda(X = ass.rel.t2, Y = init, Z = block) Model signif. = .
Total 0.534 1 R2 = 0.1095098
Conditioned 0.29726 0.5567 Variance explained by ‘block’
Constrained 0.06081 0.1139 Variance explained by ’Treatment
Unconstrained 0.17593 0.3295 Unexplained variance

Composition dissimilarity * 2020 - 2022 *

Partitioning_of_variance Inertia Proportion Note
Call:rda(X = ass.rel.t012, Y = init, Z = block) Model signif. = ***
Total 0.56175 1 R2 = 0.05628909
Conditioned 0.249 0.4433 Variance explained by ‘block’
Constrained 0.03416 0.0608 Variance explained by ’Treatment
Unconstrained 0.2786 0.4959 Unexplained variance

Q2 Why are plant communities in fallen log patches different from patches in the open?

Overview of results

  1. Soil C is significantly higher in log patches than open patches.
  • Total soil carbon is significantly higher in log patches than open patches (Wilcoxon rank sum test, p<0.05).
  1. Plant abundance has a variable relationship with different nutrient elements:
  • In 2020, we found more plants in C rich area.
  • In 2021, we found more plants in P rich area.
  • In 2022, plants abundance tend to drop with N but increase in C and base cations.
  • Overall (2020-2022), N tends to have a negative relationship with plant abundance.

  1. Relationship between plant diversity and soil elements are variable.

  2. There are no significant relationships found between plant composition and soil elements.

2020-2022 There are three response variables: count (Poisson), Shannon diversity index (hurdle: binomial and Gaussian), and composition matrix. 1. Abundance – the inclusive model includes all nutrient elements as additive fixed factors with block and year as random terms. As suggested by this inclusive model, the abundance of plants is significantly more likely to increase with soil carbon (C: p=0.004) level but decrease with plant-available nitrogen (NH4N: p<0.001) and basic exchangeable cations (CEC: p=0.008). [This seems to be a bit confounding as it does not make sense biologically. Maybe an artefact of collinearity?] 2. Diversity – the probability of having a positive Shannon diversity index in the plant community is significantly higher with higher levels of plant-available nitrogen (NO3N: p < 0.05) and potassium (K: p < 0.05) but lower with higher soil phosphorus (P: p <0.05). A similar pattern is observed for the level of diversity. The presented model has a structure of “diversity ~ elements (additive) + (1|block) + (1|year)”. 3. Composition – The composition of plant communities is not significantly explained by any of the nutrient elements. See print(en.nutrient).

2020 There are three response variables: count (Poisson), Shannon diversity index (hurdle: binomial and Gaussian), and composition matrix. 1. Abundance – the inclusive model includes all nutrient elements as additive fixed factors with block and year as random terms. As suggested by this inclusive model, the abundance of plants is significantly more likely to increase with soil carbon (C: p=0.04) level. 2. Diversity – the probability of having a positive Shannon diversity index in the plant community is significantly higher with higher levels of plant-available nitrogen (NO3N: p < 0.05) and potassium (K: p < 0.05). A similar pattern is observed for the level of diversity. The presented model has a structure of “diversity ~ elements (additive) + (1|block)”. 3. Composition – The composition of plant communities is not significantly explained by any of the nutrient elements. See print(en.nutrient).

2021 There are three response variables: count (Poisson), Shannon diversity index (hurdle: binomial and Gaussian), and composition matrix. 1. Abundance – the inclusive model includes all nutrient elements as additive fixed factors with block and year as random terms. As suggested by this inclusive model, the abundance of plants is significantly more likely to increase with plant-available phosphorus (P: p=0.03) level but decrease with plant-available nitrogen (NH4N: p=0.009). [Same comment] 2. Diversity – the level of plant diversity as represented by the Shannon diversity index and the probability of having a positive Shannon diversity index in the plant community are not significantly explained by any soil nutrient elements. The presented model has a structure of “diversity ~ elements (additive) + (1|block)”. 3. Composition – The composition of plant communities is not significantly explained by any of the nutrient elements. See print(en.nutrient).

2022 There are three response variables: count (Poisson), Shannon diversity index (hurdle: binomial and Gaussian), and composition matrix. 1. Abundance – the inclusive model includes all nutrient elements as additive fixed factors with block and year as random terms. As suggested by this inclusive model, the abundance of plants is significantly more likely to decrease with plant-available nitrogen (NH4N: p<0.001) and potassium (K: p=0.03). [Same comment] 2. Diversity – the level of plant diversity as represented by the Shannon diversity index and the probability of having a positive Shannon diversity index in the plant community are not significantly explained by any soil nutrient elements. The presented model has a structure of “diversity ~ elements (additive) + (1|block)”. 3. Composition – The composition of plant communities is not significantly explained by any of the nutrient elements. See print(en.nutrient).

Statistical Methods

  1. We performed a principal component analysis with correlation matrix on seven soil properties to understand the multivariate relationships among them.
  2. We used the Wilcoxon rank sum test to compare soil nutrient levels between log and open plots.
  3. Effect of soil nutrient on plant abundance.

This analysis involves three types of data: (1) plant counts of all plots in 2020 before the experiment setup differentiated by the initial treatment “log” and “open”; (2) plant counts from insitu_log and insitu_open plots in 2021 and 2022 respectively, including species with unknown identity; and (3) soil nutrient data collected from one random log plot and one random open plot in each block. (N=??)

Since independent nutrient variables were sampled at block level, this part of the analysis does not include block as the random term to avoid singularity issues. We used a generalized linear model (GLM) with a Poisson distribution structure to model the effect of different soil components on the abundance of plant individuals (data from 2020, 2021, and 2022). Additionally, we performed model dredging (using the MuMln package) to compare the best fitting model, the biologically sensible model, and the inclusive model with all nutrient elements.

To avoid pseudo-replication, when pulling count data from 2020 to 2022 together, we included time as a random term.

  1. Effect of soil nutrient on plant diversity (Shannon diversity index = SDI) This analysis involves three types of data: (1) plant count for each species of all plots in 2020 before the experiment setup differentiated by the initial treatment “log” and “open”; and (2) plant count for each species from insitu_log and insitu_open plots in 2021 and 2022 respectively. All data used for diversity comparison excludes species with unknown identity. (3) soil nutrient data. We used the Shannon diversity index to describe the plant species diversity for each transect at the plot level. Similarly, we used a hurdle model to model the effect of soil properties on plant diversity.
  1. Model the probability of zero and non-zero Shannon diversity with a logistic regression. GLM for single-year data and GLMER for 2020-2022 data with year as random terms in 2020-2022 data. (Question for group: what approach should we take to multiyear data? Aubrie thinks using glmer and year as random intercept)
  2. Model the non-zero Shannon diversity with an LM for single-year data and LMER for 2020-2022 data with year as a random term since the distribution of non-zero data has an approximal Gaussian distribution.
  3. We also performed model dredging (using the MuMln package) to compare the best-fitting model, the biologically sensible model, and the inclusive model with all nutrient elements.
  1. We performed similar NMDS analyses as described in the previous section. We then used envfit analyses from the Vegan package to fit soil components onto the nondimensional space.

Analysis

Nutrient composition comparison between log and open
  • Organic carbon in soil is significantly higher in fallen log patches. 
##   element    w    p_value
## 1       N 41.5 0.03126054
## 2       P 26.5 0.82963146
## 3       K 26.0 0.90151515
## 4       C 42.5 0.02518656
## 5      pH 29.5 0.56096176
## 6     CEC 30.0 0.53496503
Abundance ~ nutrient composition
Abundance ~ soil nutrient [2020-2022]
  Best model Inclusive model
Predictors Log-Mean std. Error CI p Log-Mean std. Error CI p
(Intercept) 1.82 0.02 1.78 – 1.86 <0.001 1.82 0.02 1.78 – 1.86 <0.001
N -0.08 0.02 -0.13 – -0.03 0.001 -0.08 0.03 -0.14 – -0.02 0.007
C 0.05 0.03 -0.02 – 0.12 0.149 0.06 0.04 -0.01 – 0.13 0.100
CEC -0.06 0.03 -0.12 – -0.00 0.044 -0.08 0.03 -0.14 – -0.01 0.024
P 0.04 0.03 -0.01 – 0.09 0.151
K -0.01 0.03 -0.08 – 0.05 0.703
Observations 391 391
R2 Nagelkerke 0.039 0.045
Abundance ~ soil nutrient [2022]
  Best model Selected inclusive model (2022)
Predictors Log-Mean std. Error CI p Log-Mean std. Error CI p
(Intercept) 2.18 0.04 2.11 – 2.26 <0.001 2.18 0.04 2.10 – 2.26 <0.001
N -0.16 0.05 -0.25 – -0.06 0.001 -0.18 0.06 -0.30 – -0.07 0.002
P -0.08 0.04 -0.17 – -0.00 0.052 -0.07 0.05 -0.16 – 0.02 0.149
C 0.19 0.07 0.04 – 0.33 0.009 0.16 0.07 0.02 – 0.29 0.021
CEC -0.15 0.06 -0.27 – -0.04 0.011 -0.13 0.06 -0.25 – -0.01 0.032
conductivity -0.08 0.05 -0.17 – 0.01 0.101
K -0.05 0.06 -0.17 – 0.06 0.336
Observations 84 84
R2 Nagelkerke 0.263 0.246
Abundance ~ soil nutrient [2021]
  Best model Inclusive model
Predictors Log-Mean std. Error CI p Log-Mean std. Error CI p
(Intercept) 2.00 0.04 1.91 – 2.08 <0.001 2.00 0.04 1.92 – 2.08 <0.001
K -0.08 0.05 -0.18 – 0.02 0.100 -0.09 0.06 -0.22 – 0.03 0.154
N -0.09 0.05 -0.19 – 0.01 0.073 -0.10 0.06 -0.23 – 0.02 0.098
P 0.11 0.04 0.03 – 0.19 0.009 0.12 0.04 0.03 – 0.20 0.009
C 0.04 0.08 -0.12 – 0.19 0.598
CEC -0.03 0.07 -0.16 – 0.10 0.664
Observations 86 86
R2 Nagelkerke 0.128 0.132
Abundance ~ soil nutrient [2020]
  Best model Inclusive model
Predictors Log-Mean std. Error CI p Log-Mean std. Error CI p
(Intercept) 1.52 0.03 1.46 – 1.58 <0.001 1.52 0.03 1.46 – 1.58 <0.001
P 0.06 0.03 -0.00 – 0.12 0.054 0.06 0.04 -0.02 – 0.13 0.153
C 0.07 0.03 0.01 – 0.13 0.034 0.10 0.05 -0.00 – 0.20 0.049
N 0.01 0.04 -0.07 – 0.10 0.732
K 0.07 0.05 -0.03 – 0.17 0.161
CEC -0.09 0.05 -0.20 – 0.01 0.078
Observations 221 221
R2 Nagelkerke 0.056 0.080
Diversity ~ nutrient composition
Shannon diversity index ~ soil nutrient[2020-2022]
  Best model
Predictors Estimates CI p
Count Model
(Intercept) 0.89 0.74 – 1.03 <0.001
N 0.09 0.03 – 0.14 0.002
P -0.05 -0.11 – 0.01 0.100
K 0.09 0.03 – 0.15 0.005
(Intercept) 0.49 0.45 – 0.52
Zero-Inflated Model
(Intercept) -21.71 -12101.57 – 12058.14 0.997
Random Effects
σ2 0.49
τ00 time 0.01
ICC 0.03
N time 3
Observations 393
Marginal R2 / Conditional R2 0.016 / 0.045 
## Random effect variances not available. Returned R2 does not account for random effects.
Shannon diversity index ~ soil nutrient[2022]
  Best model
Predictors Estimates std. Error CI p
Count Model
(Intercept) 0.99 0.05 0.89 – 1.09 <0.001
N 0.14 0.06 0.02 – 0.26 0.026
P -0.14 0.06 -0.25 – -0.04 0.009
K 0.18 0.06 0.06 – 0.30 0.004
(Intercept) 0.45 0.38 – 0.52
Zero-Inflated Model
(Intercept) -22.23 12526.09 -24572.90 – 24528.45 0.999
Observations 84
R2 conditional / R2 marginal NA / 0.028 
## Random effect variances not available. Returned R2 does not account for random effects.
Shannon diversity index ~ soil nutrient[2021]
  Best model
Predictors Estimates std. Error CI p
Count Model
(Intercept) 1.06 0.04 0.99 – 1.14 <0.001
CEC -0.07 0.04 -0.15 – 0.01 0.068
P 0.11 0.04 0.03 – 0.19 0.006
(Intercept) 0.32 0.27 – 0.38
Zero-Inflated Model
(Intercept) -2.30 0.38 -3.04 – -1.56 <0.001
Observations 86
R2 conditional / R2 marginal NA / 0.013 
## Random effect variances not available. Returned R2 does not account for random effects.
Shannon diversity index ~ soil nutrient[2020]
  Best model
Predictors Estimates std. Error CI p
Count Model
(Intercept) 0.71 0.03 0.65 – 0.78 <0.001
CEC -0.07 0.05 -0.17 – 0.02 0.110
K 0.15 0.05 0.05 – 0.24 0.003
N 0.13 0.04 0.05 – 0.21 0.001
(Intercept) 0.50 0.46 – 0.55
Zero-Inflated Model
(Intercept) -21.42 8108.10 -15913.01 – 15870.17 0.998
Observations 223
R2 conditional / R2 marginal NA / 0.015
Composition ~ nutrient composition

2020 - There is no significant correlation between nutrient elements and plant composition.

## 
## ***VECTORS
## 
##        NMDS1    NMDS2     r2 Pr(>r)
## N   -0.69591 -0.71813 0.1044  0.570
## P    1.00000 -0.00253 0.2459  0.209
## K    0.36669  0.93034 0.1745  0.325
## C   -0.83030  0.55732 0.1548  0.395
## CEC  0.11259  0.99364 0.1268  0.479
## Permutation: free
## Number of permutations: 999
## Permutation test for rda under reduced model
## Terms added sequentially (first to last)
## Permutation: free
## Number of permutations: 999
## 
## Model: rda(formula = ass.rel.t0 ~ N + P + K + C + CEC, data = blocksum[, c(90:93, 95)])
##          Df Variance      F Pr(>F)
## N         1 0.029990 0.8854  0.547
## P         1 0.031152 0.9197  0.514
## K         1 0.029570 0.8730  0.556
## C         1 0.041718 1.2317  0.238
## CEC       1 0.022834 0.6741  0.792
## Residual  8 0.270974

2021 - There is no significant correlation between nutrient elements and plant composition.

## [1] 14
## 
## ***VECTORS
## 
##        NMDS1    NMDS2     r2 Pr(>r)  
## N    0.44568 -0.89519 0.3458  0.092 .
## P   -0.86881  0.49515 0.1385  0.429  
## K   -0.75148  0.65975 0.1268  0.477  
## C    0.84914 -0.52816 0.0169  0.905  
## CEC -0.88005 -0.47488 0.0770  0.647  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## Permutation: free
## Number of permutations: 999
## Permutation test for rda under reduced model
## Terms added sequentially (first to last)
## Permutation: free
## Number of permutations: 999
## 
## Model: rda(formula = ass.rel.t1 ~ N + P + K + C + CEC, data = blocksum[, c(90:93, 95)])
##          Df Variance      F Pr(>F)
## N         1  0.04805 1.1232  0.313
## P         1  0.03496 0.8172  0.697
## K         1  0.02768 0.6469  0.904
## C         1  0.02952 0.6900  0.857
## CEC       1  0.05014 1.1719  0.252
## Residual  8  0.34227

2022 - There is no significant correlation between nutrient elements and plant composition.

## 
## ***VECTORS
## 
##        NMDS1    NMDS2     r2 Pr(>r)  
## N   -0.64159  0.76705 0.0366  0.812  
## P    0.95877 -0.28417 0.2378  0.225  
## K    0.61430  0.78907 0.0906  0.622  
## C   -0.02118  0.99978 0.3651  0.071 .
## CEC  0.33524  0.94213 0.1705  0.372  
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## Permutation: free
## Number of permutations: 999
## Permutation test for rda under reduced model
## Terms added sequentially (first to last)
## Permutation: free
## Number of permutations: 999
## 
## Model: rda(formula = ass.rel.t2 ~ N + P + K + C + CEC, data = blocksum[, c(90:93, 95)])
##          Df Variance      F Pr(>F)
## N         1  0.05676 1.3041  0.173
## P         1  0.03193 0.7335  0.772
## K         1  0.04159 0.9555  0.510
## C         1  0.03231 0.7423  0.800
## CEC       1  0.02320 0.5330  0.960
## Residual  8  0.34821

2020-2022 - P and pH seem to be significantly affecting the compositions of log and open communities - I doubt it since envfit function do not take ‘year’ as a random effect.

## [1] 42
## 
## ***VECTORS
## 
##        NMDS1    NMDS2     r2 Pr(>r)   
## N    0.63130 -0.77554 0.0902  0.155   
## P   -0.52734  0.84965 0.2410  0.006 **
## K   -0.45779  0.88906 0.0782  0.220   
## C    0.63281 -0.77431 0.0730  0.237   
## CEC -0.86036  0.50969 0.0032  0.939   
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## Permutation: free
## Number of permutations: 999
## Permutation test for rda under reduced model
## Terms added sequentially (first to last)
## Permutation: free
## Number of permutations: 999
## 
## Model: rda(formula = ass.rel.t012 ~ N + P + K + C + CEC, data = blocksum[, c(91:94, 96)])
##          Df Variance      F Pr(>F)  
## N         1  0.02444 1.8824  0.026 *
## P         1  0.02091 1.6107  0.040 *
## K         1  0.01568 1.2077  0.224  
## C         1  0.01996 1.5376  0.084 .
## CEC       1  0.01330 1.0239  0.411  
## Residual 36  0.46745                
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Interpretation

Q3 Are/how are plant species performances affected by proximity to fallen logs?

Overview of results

2021

There are three response variables: presence (binomial), count (truncated poisson), and per capita biomass (gaussian).

Presence - The best fit model for presence is a tie between between physical barrier and the additive model of physical barrier + nutrient island (deltaAIC=0.32). I present the model analysis of the additive model. - Both TRCY and TROR are significantly more likely to be present where there is a physical barrier as compared to in the open (TRCY: p=0.03, TROR: p=0.01) - Nutrient island does not significantly explain variation in presence/absence in any species. However, there is a trend across all species that they are more likely present in log-legacy plots.

Count - The best fit model for count is nutrient island. - Nutrient island does not strongly significantly explain variation in presence/absence in any species. However, TRCY is marginally significantly affected by nutrient island, where there are more TRCY in plots where there is a `log legacy’ (there had been a decomposing log in the plot before the experiment began)

Biomass - The best fit model for per capita biomass is the nutrient island, then physical barrier island, then the additive model. They are all within 2 AICc points of each other (deltaAIC=1.55 and 1.76, respectively). I present the model analysis of the additive model. - Nutrient island significantly explains variation GORO biomass, where there are larger GORO plants in plots where there is a `log legacy’ (p=0.03). - Physical barrier does not explain variation in biomass for any species.

2022

There are again three response variables: presence (binomial), count (truncated poisson), and per capita biomass (gaussian).

Presence 1. The best fit model for presence is a tie between between physical barrier and the additive model of physical barrier + nutrient island (deltaAIC=0.24). I present the model analysis of the additive model.

  1. Both TRCY and TROR are significantly more likely to be present where there is a physical barrier as compared to in the open (TRCY: p=0.007, TROR: p=0.03)

  2. Nutrient island significantly explains the presence/absence of TRCY, where TRCY is more likely to be present in log legacy plots as compared to open legacy plots (p=0.02)

Count 1. The best fit model for count is a tie between nutrient island and physical barrier. I present the model analysis of the nutrient island model.

  1. Nutrient island does not strongly significantly explain variation in presence/absence in any species.

Biomass 1. The best fit model for per capita biomass is the nutrient island, then the additive model, then the physical barrier model. They are all within 1 AICc points of each other (deltaAIC=0.31 and 0.46, respectively). I present the model analysis of the additive model.

  1. Nutrient island significantly explains variation in TROR biomass, where log legacy plots have larger TROR as compared to open legacy plots (p=0.04)

  2. Physical barrier significantly explains variation in TROR biomass, where plots in the open have larger TROR than plots near a physical barrier (either log or PVC pipe) (p=0.02)

Statistical Methods

The basic approach is to analyse count and biomass data from 2021 and 2022 sowing experiment. We sowed 15 seeds for each species into 16 plots in each of the seven blocks. There are 6 plot type treatments. The gap and open treatments each have four replicates per block, and the insitu_log, insitu_pvc, open_with_log, and open_with_pvc each have two replicates per block. note: For each treatment of plot type, there is also a dispersal treatment, but I do not analyse that here (yet).
- For count data I use a glmmTMB to run a generalized linear mixed effects model approach to analyse the data. The hurdle model approach I am using is to first code presence/absence as 0 or 1 (1 being any nonzero count value) and run this analysis as a binomial regression. Depending on the model fit and residual dispersion (using DHARMa), I then either run a truncated poisson or truncated negative binomial regression on the count data. Though I did run alternative versions of hurdle models that predict count while accounting for zero inflation, I felt that analyzing the presence/absence, and then analyzing count, may be revealing more of the biology of the system, where presence/absence was distinctly affected by treatments while count was less so.
- For biomass data I use linear mixed effects model to analyse the per capita biomass data. I first do a log(1+n) transformation on the per capita biomass, then analyze. I chose to do per capita biomass because of recalcitrant (!!) residual dispersion in total biomass, even after attemtps at log and square root transformations.
I chose to do a model selection approach, which I’m trying to move away from generally. However, the models I use to analyze the data represent different hypotheses about why plant species might perform differently, and so I chose to do a model comparison and selection approach as a way to not only understand how plant species performance is affected by proximity to fallen logs, but also gain insight as to why this may be the case.

N.B.:Model comparison might not be the approach we want, and we may want to run the analysis differently. I’m open to discussion and change on this point.

After model comparison, I use estimated marginal means (package emmeans) on the best fit model and compute the significance of the difference in estimates.

The models I use test the hypotheses from above: Log decomposition creates islands of fertility directly around the fallen log and fallen logs alter the microclimate directly around them by providing shade.

Analysis

2021
ggarrange(pl1, pl2, pl3, ncol=1, common.legend = T)

2022
ggarrange(pl4, pl5, pl6, ncol=1, common.legend = T)